Foreword
Artificial intelligence is one of the most broad concept, the purpose is to let the computer artificial intelligence machine can think like a man, and machine learning is a branch of artificial intelligence, specializing in computer simulation or how to achieve human learning behavior, in order to obtain a new knowledge or skills, so as to continuously improve their performance.
Deep learning is a method of machine learning, it attempts to use a complex structure comprising one or more treated layer (neural network) composed of multiple nonlinear transform algorithm for data abstraction level.
The purpose is to establish the depth of learning and simulation of the human brain to analyze learning neural network, which mimics the mechanism of the brain to interpret the data.
This is a learning knowledge points:
In deep learning network, each node in the output layer is a layer based on the previous learning to identify a specific set of features. With the increase of the depth of the neural network, the nodes can identify characteristics is more and more complicated, because each layer will feature integration and restructuring of the previous layer.
Depth learning process is equally divided into training and reasoning (both "Evaluation") two processes, data models to get through the training process, and then used to assess the new data.
Data model is divided into two types, one is the so-called discriminant model (Discriminative Model), it means that the model can be directly used to determine things. Here's judgment of things, the most typical is to do the classification. Now that can be used to direct classification, which means we can under the conditions known attributes, the record to judge. Therefore, the discriminant model is modeled on condition probability, that is, p (Y | X). Where X is the set of properties, it is actually a vector; and Y may be a value (in this case corresponds to classification), may be a vector (in this case a corresponding sequence Dimensioning). Discriminant model commonly used to deal with classification problems (such as identification of spam), image recognition and so on.
Then talk generated model (the Generative
Model). We may describe the process of generating the model data. In other words, this model is known, we can generate the data model description. The data consists of two parts, i.e. (X, Y), wherein the former is the latter category is (Y is a scalar quantity) or the sequence classification (Y is a vector). To describe the entire data, that is, to on p (X, Y) modeling, it is the joint probability modeling. Themselves do not generate a model category or marked sequence, but can be used to solve these problems, the formula may also be used to generate problems such as bot AI compose such problems.
And machine learning can be divided into several categories below:
Supervised learning from given training data set out a learning function, when the arrival of new data, the results can be predicted based on this function. Supervised learning the training set is required, including input and output, it can be said characteristics and objectives. The goal of the training data is marked by the people. Common supervised learning algorithms including regression analysis and statistical classification.
Unsupervised learning and supervised learning compared to the results of the training set is not man-made labels. Common unsupervised learning clustering algorithm.
Semi-supervised learning ranged between supervised learning and unsupervised learning. It is mainly to consider how to use a small amount of labeled samples and a large number of unlabeled samples for training and classification problems. .
Enhance learning how to learn to make action by observation. Each action will have an impact on the environment, learning objects to make judgments based on the feedback observed the surroundings.
In the traditional machine learning, supervised learning biggest problem is the relatively high cost of training data annotation, learning without supervision limited range of applications. Use a small amount of training samples and a large number of unlabeled data semi-supervised learning has been the focus of research in machine learning.
The current idea is very popular depth study GAN model and semi-supervised learning have in common, GAN is "to generate confrontation Network" (Generative Adversarial Networks) referred to, including the GAN's objective function of a generator model G and a discriminant model D, It is a zero-sum game on D and G, as well as a minimum - maximization problem.
GAN is actually generate an imitation game between the model and the discriminant model. The purpose of the model is generated, is to try to imitate, modeling and studying the distribution of real data; and discriminant model is to determine an input their own data obtained, whether it is from the real data comes from a distributed generation model . By constantly between the two internal model of competition and thereby increase the production capacity and the ability to distinguish the two models.
Current mainstream deep learning framework:
Caffe proposed by Berkeley
TensorFlow proposed by Google (Apache 2.0)
Torch (BSD License), facebook is the main user
MXNet a relatively neutral machine learning framework (Apache 2.0), is used Amazon AWS
CNTK2 proposed by Microsoft (MIT License)
Reproduced in: https: //juejin.im/post/5d04942b518825543e32d61d